nvidia repositories are hard coded

Bug #1915361 reported by Narinder Gupta
18
This bug affects 2 people
Affects Status Importance Assigned to Milestone
Containerd Subordinate Charm
High
Unassigned

Bug Description

We are trying airgap deployment of kubernetes where we handcrafted the Nvidia repositories in the local repo. But contained fails to install Nvidia packages as it hard code the Nvidia repositories and keys. This should be configurable through config options.

Tags: seg Edit Tag help
Revision history for this message
George Kraft (cynerva) wrote :
Changed in charm-containerd:
importance: Undecided → Medium
status: New → Triaged
Revision history for this message
Ante Karamatić (ivoks) wrote :

Attached is the solution Narinder used to solve this problem.

Revision history for this message
Vern Hart (vern) wrote :

Is there a plan to incorporate this patch into the charm?

I'm doing another offline deployment for a customer that requires drivers from nvidia for gpu integration.

Felipe Reyes (freyes)
tags: added: seg
Revision history for this message
Vern Hart (vern) wrote :

Subscribing field-medium as this impacts customer deployments.

The patch attached above is not quite enough. Need to add a couple of variables to the config.yaml. Also, I had to update the hard-coded packages as the versions in the package names have advanced. Specifically, the packages in the patch are:

NVIDIA_PACKAGES = [
    'cuda-drivers',
    'nvidia-container-runtime',
    'nvidia-fabricmanager-460',
    'nvidia-utils-460',
    'linux-modules-nvidia-460-generic',
]

And cuda-drivers, which appears to be a pseudo package that points to cuda-drivers-470 but the latest linux-modules-nvidia package is linux-modules-nvidia-465-generic so you get dependency conflicts. I had to update it to:

NVIDIA_PACKAGES = [
    'cuda-drivers-465',
    'nvidia-container-runtime',
    'nvidia-fabricmanager-465',
    'nvidia-utils-465',
    'linux-modules-nvidia-465-generic',
]

The way nvidia includes the software version in the package name seems to make it cumbersome to keep this up-to-date. There's got to be an automated way to determine the latest packages to install.

Revision history for this message
Narinder Gupta (narindergupta) wrote :

option option would be make package list as part of config option might help.

George Kraft (cynerva)
Changed in charm-containerd:
importance: Medium → High
To post a comment you must log in.
This report contains Public information  Edit
Everyone can see this information.

Other bug subscribers